Internal AI for Business: Building Knowledge Bases for HR, Sales, Marketing and Finance

Human Resources

Approved policies, onboarding, offboarding, benefits, travel expenses, job descriptions, training and forms can support employee questions and checklists. Employee records, salaries, health data, recruitment and disciplinary decisions require strict separation and human authority.

Sales

Product catalogues, current pricing policies, delivery terms, warranties, proposals and case studies can help with product discovery, follow-up and draft proposals. CRM reads and writes must be separate permissions; quotations and special commitments require validation and approval.

Marketing

Brand guidance, approved product information, content pillars, campaign history and approval rules can support briefs and content repurposing. Claims, prices, customer references and external publication need review.

Finance and Accounting

Expense policies, payment procedures, close checklists, invoice guidance and reconciliation procedures can help staff prepare complete records. AI must not independently approve payments, change bank details, post accounting entries or issue tax conclusions.

IT and Operations

Service catalogues, runbooks, topology, dependencies, known errors, incident and recovery procedures can assist ticket classification and troubleshooting. Production commands, firewall changes, access grants, restarts and deletion must use narrow tools with risk-based approval.

A practical model shares the user experience, model gateway, observability and governance while separating domains such as kb-public, kb-hr-policy, kb-hr-confidential, kb-sales-product, kb-sales-account, kb-marketing-brand, kb-finance-procedure and kb-it-runbook.

Isolation can use indexes, namespaces, collections or metadata filters depending on scale and sensitivity. Authorization must occur before or during retrieval. Retrieving unauthorized content and asking the model not to reveal it is not a dependable security control.

Each document should retain owner, source, department, classification, allowed groups, effective date, expiry, version and approval state. Existing ACLs should be preserved or synchronized where possible.

Sources and ingestion

Connectors may read SharePoint, Drive, CMS, file servers, CRM, ERP, HR systems, wikis, tickets, databases and APIs. Each needs an owner, scope, credentials and failure handling. Ingestion validates file safety and format, extracts and normalizes content, prevents duplication, adds metadata and permissions, and quarantines unapproved or stale material.

Indexing and retrieval

HR policy, product tables, contracts and technical runbooks require different chunking and evaluation sets. Retrieval may combine keyword, semantic, vector or hybrid search.

Identity and policy

Users authenticate through SSO. Trusted groups, roles and attributes drive knowledge-domain and tool access. The application must fail closed when authorization cannot be evaluated.

Orchestration and model gateway

The orchestrator classifies intent, chooses domains, applies security filters, retrieves evidence and sends only authorized passages to the model. Answers cite sources and clarify, abstain or escalate when evidence is insufficient. The gateway centralizes model routing, quotas, secrets, cost attribution and rate limiting.

Tools and workflows

Use narrow operations such as crm.search_customer, crm.create_draft_note, ticket.create and invoice.validate_fields. Sensitive actions require validation, authorization, approval, idempotency, result verification and audit.

Observability and evaluation

Monitor latency, errors, departmental cost, retrieval quality, citation correctness, abstention, access denial, stale documents, tool failures and approval timeout. Retain enough evidence for investigation without unnecessary personal data, secrets or private reasoning.

HR leaders own employee policies; Product or Sales Operations owns product and pricing knowledge; Marketing owns brand guidance; Finance owns accounting procedures; IT Operations owns runbooks.

Knowledge owners approve sources and versions, define effective dates, resolve conflicts, review unanswered questions and retire obsolete content. An AI system cannot determine which of two conflicting policies is authoritative without governance.

Update pipelines must detect changes, rebuild content, invalidate stale indexes or caches and rerun regression tests. Uploading a new document does not automatically retire the previous version.

Days 1–15: Use-case and data inventory

Select two departments, collect 20–50 real questions per department, identify sources and owners, classify data and capture a search-time baseline. Explicitly exclude unsuitable data.

Days 16–30: Retrieval-only MVP

Implement SSO, domains, ingestion, metadata, access filtering, evaluation, citations and abstention. Do not enable autonomous actions. The MVP must prove correct retrieval and authorization.

Days 31–60: Real-user pilot

Invite representative users, analyze unsupported questions, measure citation quality, test permissions, improve content and retrieval, and track cost and saved effort. Do not use prompts to hide poor source quality.

Days 61–90: Governed expansion

Add the next department, standardize owner onboarding, build dashboards, add drafting workflows and only then enable small reversible actions with approval, audit, incident response and rollback.

Knowledge quality: grounded-answer rate, citation correctness, retrieval recall, stale-source usage and update latency.

Business outcomes: search time, repeated questions, onboarding time, proposal preparation and draft acceptance after review.

Risk controls: unauthorized retrieval, policy-blocked tool calls, duplicate side effects, approval timeout, data incidents and correct abstention.

Operations and cost: cost by department and use case, p50/p95 latency, error rate, ingestion failures, ownerless documents and active adoption. Tie cost to completed work rather than tokens alone.

One unrestricted store: increases leakage, stale versions and ownership ambiguity.

Treating upload as completion: knowledge requires versioning, expiry, approval and regression tests.

Using prompts as security: access control belongs in the application, retrieval and tool layers.

Connecting ERP or CRM too early: prove identity and retrieval before enabling side effects.

No knowledge owner: conflicting and obsolete documents continue to influence answers.

Measuring question volume instead of outcomes: high usage does not by itself prove productivity or quality.

Does internal AI require a custom-trained model?

Not necessarily. RAG is suitable for frequently changing knowledge. Fine-tuning is more appropriate for behavior or output format and does not replace knowledge governance.

Can one assistant serve the entire company?

The interface and platform can be shared, while knowledge domains, permissions, prompts, tools and KPIs remain use-case specific.

Can it connect to SharePoint, CRM and ERP?

Yes, but every source needs its own scope, owner, credentials and failure handling. Read access does not imply update permission.

Which department should start?

Choose repeated questions, clear sources, available ownership and moderate risk. General HR policy, Sales product knowledge or IT service-desk knowledge are usually safer than payments or employment decisions.

Sources used

• Microsoft: Retrieval-augmented generation and indexes
• Microsoft: Design and develop a RAG solution
• Microsoft: Document-level access control
• NIST: Generative AI Profile — AI RMF

Continue reading

• Internal AI assistant: Prevent unauthorized document answers
• AI workflow automation